Error analysis and modifications to the short-time speech transmission index

• Published in: The Journal of the Acoustical Society of America 2014-04, Vol.135 (4_Supplement), p.2224-2224
• Main Authors: Payton, Karen, Ferreira, Matthew
• Format: Article
• Language: English
• ISSN: 0001-4966; 1520-8524
• DOI: 10.1121/1.4877271

The Speech Transmission Index (STI) predicts the intelligibility of speech degraded by noise and reverberation. Recently, Payton and Shrestha [J. Acoust. Soc. Am. 134, 3818–3827 (2013)] reported on a short-time speech-based STI (ssSTI) to predict time-varying intelligibility of speech using analysis windows shorter than 1 s. While the ssSTI generally tracked a theoretical STI calculated using octave-band signal-to-noise ratios, it deviated from the theoretical calculation for windows shorter than 0.3 s. The current work analyzes and improves the performance of the ssSTI for speech degraded by stationary speech-shaped noise. Using a cluster analysis, the time-varying standard deviation was determined to be inversely proportional to window length, octave band and speech envelope variance. Two ssSTI modifications are proposed to reduce the 0.3 s window limitation: A silence detection algorithm eliminates non-zero ssSTI values that occur during silence and a modified envelope extraction scheme reduces the standard deviation by increasing envelope bandwidth. Using the 0.3 s window as a performance benchmark, new octave-band specific window limitations, ranging from 151 ms to 21 ms, were established. The modified ssSTI also works with common octave-band window lengths as short as 30 ms when full envelope bandwidths are used in combination with the silence detector.